Familiar configuration enables figure/ground assignment in natural scenes

Abstract

VSS05 Abstract Figure/ground organization is a step of perceptual organization that assigns a contour to one of the two abutting regions. Peterson et al showed that configurations of contours, such as outlines of recognizable objects, provide a powerful cue that can dominate traditional f/g cues such as symmetry. In this work we: (1) provide an operationalization of familiar configuration in terms of prototypical local shapes, without requiring global object recognition; (2) show that a classifier based on this cue works well on images of natural scenes. Shape context [Belongie, Malik & Puzicha ICCV01,Berg & Malik CVPR01] is a shape descriptor which summarizes local arrangement of edges, relative to the center point, in a logpolar fashion. We cluster a large set of these descriptors to construct a small list of prototypical shape configurations, or shapemes (analogous to phonemes). Shapemes capture important local structures such as convexity and parallelism. For each point along a contour, we measure the similarity of its local shape descriptor to each shapeme. These measurements are combined using a logistic regression classifier to predict the figure/ground label. We test it on a Berkeley Figure/ground Dataset which consists of 200 natural images w/ human-marked f/g labels. By averaging the classifier outputs over all points on each contour, we obtain an accuracy of 72% (chance is 50%). This compares favorably to the traditional f/g cues used in [Fowlkes et al 03]. Enforcing consistency constraints at junctions increases the accuracy further to 79%, making it a promising model of figure/ground organization. Figure/Ground Organization Berkeley Figure/Ground Dataset